text input that could be read as NA/None is now read as a string #215
Conversation
| if trainer.is_world_process_zero(): | ||
| if training_args.do_train: | ||
| trainer.save_model() | ||
| trainer.save_model() # NOTE: a RobertaConfig is loaded here. why? |
There was a problem hiding this comment.
did you want to keep this in here?
| raise NotImplementedError( | ||
| "This functionality has not been restored yet" | ||
| ) | ||
| model = CnlpModelForClassification( |
There was a problem hiding this comment.
I think this definition will load a fine-tuned cnlpt model as an encoder, but re-initialize the classifier head. This would be the expected behavior for some use cases, but missing some use cases. I think we want to edit this to explicitly handle the two different cases (even if one still throws an exception), rather than having the user guess what might be happening. We should force them to specify whether to keep or ignore existing classifiers (as in the hier model).
| return compute_metrics_fn | ||
|
|
||
| # Initialize our Trainer | ||
| training_args.load_best_model_at_end = True |
There was a problem hiding this comment.
@tmills I removed this line because it makes it impossible to do prediction without training; it looks for a model checkpoint in the output dir, but when do_train = False, that doesn't exist.
No description provided.